Early threshold calibration reduced false alarms and missed most key responses in continuous SEEG replay from eight patients

We tested a fixed SEEG decoder as a continuous response trigger. We used all eight patients in OpenNeuro ds004473, version 1.0.2. For each patient, the first 40% of valid trials trained the model. The next 20 trials set a threshold with at most one observed false alarm per non-target minute. The final 40 trials tested the fixed model and threshold. The test had 320 responses and 38.18 minutes of replay. Mean event-window balanced accuracy was 72.0%. The zero threshold gave 38.89 false alarms per minute and detected 45.6% of responses before the recorded key time. The calibrated threshold gave 1.66 false alarms per minute and detected 9.7%. Its later rate exceeded one per minute in seven of eight patients. Descriptive 95% patient bootstrap intervals were 1.09 to 2.16 for the calibrated rate and 4.4% to 16.6% for detection. A rule that also accepted short post-response alarms detected 21.9% after calibration. The package preserves four figures, every continuous score, alarm times, model weights, source hashes, raw fixtures, code, and repeat instructions. This is an offline test in one task recording per patient. The short calibration period cannot certify rare false alarms. The labels are recorded key times. The test does not establish movement intention or stimulation timing.

stereoelectroencephalography · continuous decoding · false alarms · brain-computer interfaces · threshold calibration · offline replay